A self-improving LLM agent with optimization memory raises average kernel throughput from 45-49% to 59-61% of peak on Trainium accelerators and matches proprietary models at 26x lower cost.
The nc_matmul instruction must read inputs from SBUF and write outputs to PSUM
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization
A self-improving LLM agent with optimization memory raises average kernel throughput from 45-49% to 59-61% of peak on Trainium accelerators and matches proprietary models at 26x lower cost.